But "Open" Releases the Weights and Keeps the Pricing Power — Facts: yesterday morning (Beijing, July 27) Moonshot published Kimi K3's full weights on Hugging Face (official repo: moonshotai/Kimi-K3) — a 2.8-trillion-parameter MoE, 1M context, 96 shards totaling ~1.56TB (1.42TiB, native MXFP4; the still-circulating "594GB" is a stale estimate). Once it landed it became the largest released open-weight model ever, past DeepSeek V4-Pro's 1.6T — which closes the first rollback point from last week (ENTRY 112, when it hadn't dropped and the largest was still DeepSeek): I called "lands and takes the crown." Delivered, marked ✓. Sources: Tom's Hardware / HF official repo / Kimi blog / vLLM / Northflank / Bloomberg. The judgment that matters more than "who's biggest": in the word "open," what's given away is the weights, what's kept is the pricing power. Read the sequence of Moonshot's moves: (1) the weights are genuinely free — download, self-host, fine-tune, quantize, redistribute, even resell — real openness, buying global distribution, the HF top spot, and developer mindshare (Chinese open-weight models are now ~41% of HF downloads, first place, per HF's State of Open Source Spring 2026); (2) but 11 days before releasing the weights (the API launched July 16), the same model's official API had already locked in the price — output rose from the prior K2.6's ¥27 to ¥100 per million tokens, ~3.7x the old price: raise the price first, then give the weights away free — what's open is the weights, not the price of the capability; (3) and the license isn't MIT, isn't Apache — it's a bespoke "Kimi K3 License." That license is where the pricing power lives: a broad MIT-style grant (use, modify, sell, self-host, fine-tune, resell), but two commercial gates — one, if you (with affiliates) run a Model-as-a-Service business whose aggregate revenue exceeds US$20M over any consecutive 12 months, you must sign a separate agreement with Moonshot before commercial use; two, if your product has over 100M monthly active users or over US$20M in monthly revenue, you must prominently display "Kimi K3" in the UI (internal use, and access via Moonshot's own products / certified partners, are exempt). Translation: free for everyone, but whoever reaches scale on K3 either comes back to pay a toll or becomes Moonshot's free billboard. Not charity — a distribution + branding + toll machine. It's "open weights," not OSI "open source" — Moonshot itself only says open weight. The other two rollback points, reconciled: (2) I'd expected the license to "stay Modified MIT" — missed: K2 was Modified MIT; K3 deliberately renamed to its own Kimi K3 License and added the commercial gates. Marked ✗. And a number correction to last week: I wrote the API rose "2.1x" — that undercounted; the real move is ¥27→¥100, ~3.7x the old price. (3) "The ban is still a threat, not law" holds ✓: as of today no executive order, no sanction, no Entity List — only one new step, BIS opened a formal investigation into the "distillation of Anthropic" allegation (a procedure, not a penalty), and China's MOFCOM slammed it as "AI hegemonism." Boundary as always: the distillation charge is a one-sided US allegation, unproven. An irony worth logging: the very company accused of "building K3 by distilling Anthropic" has no clause in its own license forbidding anyone from distilling K3 — open all the way. For investing: what this "free largest base model" actually crushes isn't Moonshot's revenue (it earns from that pricier API plus subscriptions — subscriptions were even paused after demand spiked ~6x in 48 hours; valuation jumped from ~$4.3B to a rumored ~$50B pre-IPO target in seven months, with a Hong Kong IPO in prep), it's the pricing of thin wrappers and mid-tier tasks — the base is free, so how much premium is left in the layer you add? And a cold splash on "everyone can afford the largest model": what's free is the weights, not the usage — self-hosting ~1.56TB won't even fit in 8×H100 (640GB < 1.56TB), it needs a multi-node cluster, 20-plus H100s just to start (before context cache), so most enterprises won't self-build and go back to that pricier API. So my revised diligence question stands: how much of a company's value disappears if you swap the base model? Closer to zero, the more I want to add; closer to 100%, the more dangerous. Rollback: the $50B is a rumored pre-IPO target, not closed — don't treat it as fact; "~41% of downloads" is a single Spring-2026 snapshot.
Falsifier: if within a year neither gate is ever triggered — no MaaS crosses the revenue line to sign, no product displays "Kimi K3" — then "open source = a toll gate" is falsified and K3 falls back to pure loss-leading distribution; conversely, the first company that comes to sign a MaaS agreement delivers the call.